def main(unused_argv): vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, image_features, indicator, word_ids, pointer_ids = import_mscoco( mode="train", batch_size=FLAGS.batch_size, num_epochs=FLAGS.num_epochs, is_mini=FLAGS.is_mini) lengths = tf.reduce_sum(indicator, [1]) show_and_tell_cell = ShowAndTellCell(300) best_first_image_captioner = BestFirstImageCaptioner(show_and_tell_cell, vocab, pretrained_matrix) word_logits, wids, pointer_logits, pids, ids, _lengths = best_first_image_captioner( mean_image_features=image_features, word_ids=word_ids, pointer_ids=pointer_ids, lengths=lengths) tf.losses.sparse_softmax_cross_entropy(pointer_ids, pointer_logits) tf.losses.sparse_softmax_cross_entropy(word_ids, word_logits) loss = tf.losses.get_total_loss() global_step = tf.train.get_or_create_global_step() optimizer = tf.train.AdamOptimizer() learning_step = optimizer.minimize(loss, var_list=best_first_image_captioner.variables, global_step=global_step) captioner_saver = tf.train.Saver(var_list=best_first_image_captioner.variables + [global_step]) captioner_ckpt, captioner_ckpt_name = get_best_first_checkpoint() with tf.Session() as sess: sess.run(tf.variables_initializer(optimizer.variables())) if captioner_ckpt is not None: captioner_saver.restore(sess, captioner_ckpt) else: sess.run(tf.variables_initializer(best_first_image_captioner.variables + [global_step])) captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = time.time() for i in itertools.count(): time_start = time.time() try: twids, tpids, _ids, _lengths, _loss, _learning_step = sess.run([ word_ids, pointer_ids, ids, lengths, loss, learning_step]) except: break iteration = sess.run(global_step) insertion_sequence = insertion_sequence_to_array(twids, tpids, _lengths, vocab) print(PRINT_STRING.format( iteration, _loss, list_of_ids_to_string(insertion_sequence[0], vocab), list_of_ids_to_string(twids[0, :].tolist(), vocab), FLAGS.batch_size / (time.time() - time_start))) new_save = time.time() if new_save - last_save > 3600: # save the model every hour captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = new_save captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) print("Finishing training.")
PRINT_STRING = """({3:.2f} img/sec) iteration: {0:05d}\n caption: {1}\n label: {2}""" BATCH_SIZE = 10 BEAM_SIZE = 16 if __name__ == "__main__": vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, mean_features, input_seq, target_seq, indicator = ( import_mscoco(mode="train", batch_size=BATCH_SIZE, num_epochs=1, is_mini=True)) image_captioner = ImageCaptioner(ShowAndTellCell(300), vocab, pretrained_matrix, trainable=False, beam_size=BEAM_SIZE) logits, ids = image_captioner(mean_image_features=mean_features) captioner_saver = tf.train.Saver( var_list=remap_decoder_name_scope(image_captioner.variables)) captioner_ckpt, captioner_ckpt_name = get_show_and_tell_checkpoint() with tf.Session() as sess: assert (captioner_ckpt is not None) captioner_saver.restore(sess, captioner_ckpt) used_ids = set() json_dump = []
def main(unused_argv): vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, mean_features, input_seq, target_seq, indicator = ( import_mscoco(mode="train", batch_size=BATCH_SIZE, num_epochs=100, is_mini=True)) show_and_tell_cell = ShowAndTellCell(300) image_captioner = ImageCaptioner(show_and_tell_cell, vocab, pretrained_matrix) logits, ids = image_captioner(lengths=tf.reduce_sum(indicator, axis=1), mean_image_features=mean_features, seq_inputs=input_seq) tf.losses.sparse_softmax_cross_entropy(target_seq, logits, weights=indicator) loss = tf.losses.get_total_loss() global_step = tf.train.get_or_create_global_step() learning_rate = tf.train.exponential_decay( INITIAL_LEARNING_RATE, global_step, (TRAINING_EXAMPLES // BATCH_SIZE) * EPOCHS_PER_DECAY, DECAY_RATE, staircase=True) learning_step = tf.train.GradientDescentOptimizer( learning_rate).minimize(loss, var_list=image_captioner.variables, global_step=global_step) captioner_saver = tf.train.Saver(var_list=image_captioner.variables + [global_step]) captioner_ckpt, captioner_ckpt_name = get_show_and_tell_checkpoint() with tf.Session() as sess: if captioner_ckpt is not None: captioner_saver.restore(sess, captioner_ckpt) else: sess.run( tf.variables_initializer(image_captioner.variables + [global_step])) captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = time.time() for i in itertools.count(): time_start = time.time() try: _ids, _loss, _learning_step = sess.run( [ids, loss, learning_step]) except: break iteration = sess.run(global_step) print( PRINT_STRING.format( iteration, _loss, list_of_ids_to_string(_ids[0, :].tolist(), vocab), BATCH_SIZE / (time.time() - time_start))) new_save = time.time() if new_save - last_save > 3600: # save the model every hour captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = new_save captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) print("Finishing training.")
tf.flags.DEFINE_boolean("is_mini", False, "") tf.flags.DEFINE_string("mode", "eval", "") FLAGS = tf.flags.FLAGS if __name__ == "__main__": vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, image_features, indicator, word_ids, pointer_ids = import_mscoco( mode=FLAGS.mode, batch_size=FLAGS.batch_size, num_epochs=1, is_mini=FLAGS.is_mini) image_captioner = BestFirstImageCaptioner(ShowAndTellCell(300), vocab, pretrained_matrix, trainable=False, beam_size=FLAGS.beam_size) word_logits, wids, pointer_logits, pids, ids, _lengths = image_captioner( mean_image_features=image_features) captioner_saver = tf.train.Saver( var_list=remap_decoder_name_scope(image_captioner.variables)) captioner_ckpt, captioner_ckpt_name = get_best_first_checkpoint() with tf.Session() as sess: assert (captioner_ckpt is not None) captioner_saver.restore(sess, captioner_ckpt) used_ids = set()
def main(unused_argv): vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) attribute_map, attribute_embeddings_map = get_visual_attributes( ), np.random.normal(0, 0.1, [1000, 2048]) with tf.Graph().as_default(): image_id, mean_features, input_seq, target_seq, indicator = import_mscoco( mode="train", batch_size=FLAGS.batch_size, num_epochs=FLAGS.num_epochs, is_mini=FLAGS.is_mini) show_and_tell_cell = ShowAndTellCell(300, num_image_features=4096) attribute_image_captioner = AttributeImageCaptioner( show_and_tell_cell, vocab, pretrained_matrix, attribute_map, attribute_embeddings_map) attribute_detector = AttributeDetector(1000) _, top_k_attributes = attribute_detector(mean_features) logits, ids = attribute_image_captioner( top_k_attributes, lengths=tf.reduce_sum(indicator, axis=1), mean_image_features=mean_features, seq_inputs=input_seq) tf.losses.sparse_softmax_cross_entropy(target_seq, logits, weights=indicator) loss = tf.losses.get_total_loss() global_step = tf.train.get_or_create_global_step() optimizer = tf.train.AdamOptimizer() learning_step = optimizer.minimize( loss, var_list=attribute_image_captioner.variables, global_step=global_step) captioner_saver = tf.train.Saver( var_list=attribute_image_captioner.variables + [global_step]) attribute_detector_saver = tf.train.Saver( var_list=attribute_detector.variables) captioner_ckpt, captioner_ckpt_name = get_show_and_tell_attribute_checkpoint( ) attribute_detector_ckpt, attribute_detector_ckpt_name = get_attribute_detector_checkpoint( ) with tf.Session() as sess: sess.run(tf.variables_initializer(optimizer.variables())) if captioner_ckpt is not None: captioner_saver.restore(sess, captioner_ckpt) else: sess.run( tf.variables_initializer( attribute_image_captioner.variables + [global_step])) if attribute_detector_ckpt is not None: attribute_detector_saver.restore(sess, attribute_detector_ckpt) else: sess.run(tf.variables_initializer( attribute_detector.variables)) captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = time.time() for i in itertools.count(): time_start = time.time() try: _target, _ids, _loss, _learning_step = sess.run( [target_seq, ids, loss, learning_step]) except: break iteration = sess.run(global_step) print( PRINT_STRING.format( iteration, _loss, list_of_ids_to_string(_ids[0, :].tolist(), vocab), list_of_ids_to_string(_target[0, :].tolist(), vocab), FLAGS.batch_size / (time.time() - time_start))) new_save = time.time() if new_save - last_save > 3600: # save the model every hour captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) last_save = new_save captioner_saver.save(sess, captioner_ckpt_name, global_step=global_step) print("Finishing training.")
tf.logging.set_verbosity(tf.logging.INFO) tf.flags.DEFINE_integer("batch_size", 1, "") tf.flags.DEFINE_integer("beam_size", 3, "") tf.flags.DEFINE_boolean("is_mini", False, "") tf.flags.DEFINE_string("mode", "eval", "") FLAGS = tf.flags.FLAGS if __name__ == "__main__": vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) with tf.Graph().as_default(): image_id, mean_features, input_seq, target_seq, indicator = import_mscoco( mode=FLAGS.mode, batch_size=FLAGS.batch_size, num_epochs=1, is_mini=FLAGS.is_mini) image_captioner = ImageCaptioner(ShowAndTellCell(300), vocab, pretrained_matrix, trainable=False, beam_size=FLAGS.beam_size) logits, ids = image_captioner(mean_image_features=mean_features) captioner_saver = tf.train.Saver(var_list=remap_decoder_name_scope(image_captioner.variables)) captioner_ckpt, captioner_ckpt_name = get_show_and_tell_checkpoint() with tf.Session() as sess: assert(captioner_ckpt is not None) captioner_saver.restore(sess, captioner_ckpt) used_ids = set() json_dump = [] for i in itertools.count(): time_start = time.time() try:
FLAGS = tf.flags.FLAGS if __name__ == "__main__": vocab, pretrained_matrix = load_glove(vocab_size=100000, embedding_size=300) attribute_map, attribute_embeddings_map = get_visual_attributes( ), np.random.normal(0, 0.1, [1000, 2048]) with tf.Graph().as_default(): image_id, mean_features, input_seq, target_seq, indicator = import_mscoco( mode=FLAGS.mode, batch_size=FLAGS.batch_size, num_epochs=1, is_mini=FLAGS.is_mini) show_and_tell_cell = ShowAndTellCell(300, num_image_features=4096) attribute_image_captioner = AttributeImageCaptioner( show_and_tell_cell, vocab, pretrained_matrix, attribute_map, attribute_embeddings_map) attribute_detector = AttributeDetector(1000) _, top_k_attributes = attribute_detector(mean_features) logits, ids = attribute_image_captioner( top_k_attributes, mean_image_features=mean_features) captioner_saver = tf.train.Saver(var_list=remap_decoder_name_scope( attribute_image_captioner.variables)) attribute_detector_saver = tf.train.Saver( var_list=attribute_detector.variables) captioner_ckpt, captioner_ckpt_name = get_show_and_tell_attribute_checkpoint( ) attribute_detector_ckpt, attribute_detector_ckpt_name = get_attribute_detector_checkpoint(